K-MEANS-BASED ANOMALY DETECTION FOR GAS TURBINE ENGINE PREDICTIVE MAINTENANCE

Abstract

Gas turbine engines are important assets in power generation and industrial operations, where unexpected faults can result in equipment damage, unplanned downtime, production losses, and increased maintenance costs. Predictive maintenance provides an opportunity to identify abnormal operating behaviour before it develops into a more serious failure. However, the availability of accurately labelled fault data remains a challenge in many industrial applications. This study presents the application of K-means clustering as an unsupervised machine-learning approach for identifying normal and anomalous operating conditions in gas turbine engine operational data. The study forms part of a broader predictive fault maintenance model developed for a gas turbine engine. Gas turbine operational data were obtained from a publicly available dataset and subjected to preprocessing and normalization before clustering. The operational variables considered include ambient temperature, ambient pressure, ambient humidity, air filter differential pressure, gas turbine exhaust pressure, turbine inlet temperature, turbine after temperature, turbine energy yield, and compressor discharge pressure. Each observation was represented as a sensor measurement vector, and K-means was applied to group observations with similar operating characteristics. Euclidean distance was used to determine the proximity of each observation to the cluster centroids, while the resulting cluster characteristics were used to distinguish normal from anomalous operational patterns. Unlike supervised classification, the approach does not require a pre-existing fault-class label for every observation. The resulting cluster labels provide an operational-state representation that can subsequently support temporal fault prediction in the wider predictive maintenance model. The study demonstrates how unsupervised clustering can provide an effective preliminary anomaly-detection stage for data-driven gas turbine predictive maintenance.